In this paper, we introduce the new task of reconstructing 3D human pose froma single image in which we can see the person and the person's image through amirror. Compared to general scenarios of 3D pose estimation from a single view,the mirror reflection provides an additional view for resolving the depthambiguity. We develop an optimization-based approach that exploits mirrorsymmetry constraints for accurate 3D pose reconstruction. We also provide amethod to estimate the surface normal of the mirror from vanishing points inthe single image. To validate the proposed approach, we collect a large-scaledataset named Mirrored-Human, which covers a large variety of human subjects,poses and backgrounds. The experiments demonstrate that, when trained onMirrored-Human with our reconstructed 3D poses as pseudo ground-truth, theaccuracy and generalizability of existing single-view 3D pose estimators can belargely improved.